1// SPDX-License-Identifier: GPL-2.0+
2/*
3 * Read-Copy Update mechanism for mutual exclusion, the Bloatwatch edition.
4 *
5 * Copyright IBM Corporation, 2008
6 *
7 * Author: Paul E. McKenney <paulmck@linux.ibm.com>
8 *
9 * For detailed explanation of Read-Copy Update mechanism see -
10 *		Documentation/RCU
11 */
12#include <linux/completion.h>
13#include <linux/interrupt.h>
14#include <linux/notifier.h>
15#include <linux/rcupdate_wait.h>
16#include <linux/kernel.h>
17#include <linux/export.h>
18#include <linux/mutex.h>
19#include <linux/sched.h>
20#include <linux/types.h>
21#include <linux/init.h>
22#include <linux/time.h>
23#include <linux/cpu.h>
24#include <linux/prefetch.h>
25#include <linux/slab.h>
26#include <linux/mm.h>
27
28#include "rcu.h"
29
30/* Global control variables for rcupdate callback mechanism. */
31struct rcu_ctrlblk {
32	struct rcu_head *rcucblist;	/* List of pending callbacks (CBs). */
33	struct rcu_head **donetail;	/* ->next pointer of last "done" CB. */
34	struct rcu_head **curtail;	/* ->next pointer of last CB. */
35	unsigned long gp_seq;		/* Grace-period counter. */
36};
37
38/* Definition for rcupdate control block. */
39static struct rcu_ctrlblk rcu_ctrlblk = {
40	.donetail	= &rcu_ctrlblk.rcucblist,
41	.curtail	= &rcu_ctrlblk.rcucblist,
42	.gp_seq		= 0 - 300UL,
43};
44
45void rcu_barrier(void)
46{
47	wait_rcu_gp(call_rcu_hurry);
48}
49EXPORT_SYMBOL(rcu_barrier);
50
51/* Record an rcu quiescent state.  */
52void rcu_qs(void)
53{
54	unsigned long flags;
55
56	local_irq_save(flags);
57	if (rcu_ctrlblk.donetail != rcu_ctrlblk.curtail) {
58		rcu_ctrlblk.donetail = rcu_ctrlblk.curtail;
59		raise_softirq_irqoff(RCU_SOFTIRQ);
60	}
61	WRITE_ONCE(rcu_ctrlblk.gp_seq, rcu_ctrlblk.gp_seq + 2);
62	local_irq_restore(flags);
63}
64
65/*
66 * Check to see if the scheduling-clock interrupt came from an extended
67 * quiescent state, and, if so, tell RCU about it.  This function must
68 * be called from hardirq context.  It is normally called from the
69 * scheduling-clock interrupt.
70 */
71void rcu_sched_clock_irq(int user)
72{
73	if (user) {
74		rcu_qs();
75	} else if (rcu_ctrlblk.donetail != rcu_ctrlblk.curtail) {
76		set_tsk_need_resched(current);
77		set_preempt_need_resched();
78	}
79}
80
81/*
82 * Reclaim the specified callback, either by invoking it for non-kfree cases or
83 * freeing it directly (for kfree). Return true if kfreeing, false otherwise.
84 */
85static inline bool rcu_reclaim_tiny(struct rcu_head *head)
86{
87	rcu_callback_t f;
88	unsigned long offset = (unsigned long)head->func;
89
90	rcu_lock_acquire(&rcu_callback_map);
91	if (__is_kvfree_rcu_offset(offset)) {
92		trace_rcu_invoke_kvfree_callback("", head, offset);
93		kvfree((void *)head - offset);
94		rcu_lock_release(&rcu_callback_map);
95		return true;
96	}
97
98	trace_rcu_invoke_callback("", head);
99	f = head->func;
100	debug_rcu_head_callback(head);
101	WRITE_ONCE(head->func, (rcu_callback_t)0L);
102	f(head);
103	rcu_lock_release(&rcu_callback_map);
104	return false;
105}
106
107/* Invoke the RCU callbacks whose grace period has elapsed.  */
108static __latent_entropy void rcu_process_callbacks(struct softirq_action *unused)
109{
110	struct rcu_head *next, *list;
111	unsigned long flags;
112
113	/* Move the ready-to-invoke callbacks to a local list. */
114	local_irq_save(flags);
115	if (rcu_ctrlblk.donetail == &rcu_ctrlblk.rcucblist) {
116		/* No callbacks ready, so just leave. */
117		local_irq_restore(flags);
118		return;
119	}
120	list = rcu_ctrlblk.rcucblist;
121	rcu_ctrlblk.rcucblist = *rcu_ctrlblk.donetail;
122	*rcu_ctrlblk.donetail = NULL;
123	if (rcu_ctrlblk.curtail == rcu_ctrlblk.donetail)
124		rcu_ctrlblk.curtail = &rcu_ctrlblk.rcucblist;
125	rcu_ctrlblk.donetail = &rcu_ctrlblk.rcucblist;
126	local_irq_restore(flags);
127
128	/* Invoke the callbacks on the local list. */
129	while (list) {
130		next = list->next;
131		prefetch(next);
132		debug_rcu_head_unqueue(list);
133		local_bh_disable();
134		rcu_reclaim_tiny(list);
135		local_bh_enable();
136		list = next;
137	}
138}
139
140/*
141 * Wait for a grace period to elapse.  But it is illegal to invoke
142 * synchronize_rcu() from within an RCU read-side critical section.
143 * Therefore, any legal call to synchronize_rcu() is a quiescent state,
144 * and so on a UP system, synchronize_rcu() need do nothing, other than
145 * let the polled APIs know that another grace period elapsed.
146 *
147 * (But Lai Jiangshan points out the benefits of doing might_sleep()
148 * to reduce latency.)
149 *
150 * Cool, huh?  (Due to Josh Triplett.)
151 */
152void synchronize_rcu(void)
153{
154	RCU_LOCKDEP_WARN(lock_is_held(&rcu_bh_lock_map) ||
155			 lock_is_held(&rcu_lock_map) ||
156			 lock_is_held(&rcu_sched_lock_map),
157			 "Illegal synchronize_rcu() in RCU read-side critical section");
158	WRITE_ONCE(rcu_ctrlblk.gp_seq, rcu_ctrlblk.gp_seq + 2);
159}
160EXPORT_SYMBOL_GPL(synchronize_rcu);
161
162static void tiny_rcu_leak_callback(struct rcu_head *rhp)
163{
164}
165
166/*
167 * Post an RCU callback to be invoked after the end of an RCU grace
168 * period.  But since we have but one CPU, that would be after any
169 * quiescent state.
170 */
171void call_rcu(struct rcu_head *head, rcu_callback_t func)
172{
173	static atomic_t doublefrees;
174	unsigned long flags;
175
176	if (debug_rcu_head_queue(head)) {
177		if (atomic_inc_return(&doublefrees) < 4) {
178			pr_err("%s(): Double-freed CB %p->%pS()!!!  ", __func__, head, head->func);
179			mem_dump_obj(head);
180		}
181
182		if (!__is_kvfree_rcu_offset((unsigned long)head->func))
183			WRITE_ONCE(head->func, tiny_rcu_leak_callback);
184		return;
185	}
186
187	head->func = func;
188	head->next = NULL;
189
190	local_irq_save(flags);
191	*rcu_ctrlblk.curtail = head;
192	rcu_ctrlblk.curtail = &head->next;
193	local_irq_restore(flags);
194
195	if (unlikely(is_idle_task(current))) {
196		/* force scheduling for rcu_qs() */
197		resched_cpu(0);
198	}
199}
200EXPORT_SYMBOL_GPL(call_rcu);
201
202/*
203 * Store a grace-period-counter "cookie".  For more information,
204 * see the Tree RCU header comment.
205 */
206void get_completed_synchronize_rcu_full(struct rcu_gp_oldstate *rgosp)
207{
208	rgosp->rgos_norm = RCU_GET_STATE_COMPLETED;
209}
210EXPORT_SYMBOL_GPL(get_completed_synchronize_rcu_full);
211
212/*
213 * Return a grace-period-counter "cookie".  For more information,
214 * see the Tree RCU header comment.
215 */
216unsigned long get_state_synchronize_rcu(void)
217{
218	return READ_ONCE(rcu_ctrlblk.gp_seq);
219}
220EXPORT_SYMBOL_GPL(get_state_synchronize_rcu);
221
222/*
223 * Return a grace-period-counter "cookie" and ensure that a future grace
224 * period completes.  For more information, see the Tree RCU header comment.
225 */
226unsigned long start_poll_synchronize_rcu(void)
227{
228	unsigned long gp_seq = get_state_synchronize_rcu();
229
230	if (unlikely(is_idle_task(current))) {
231		/* force scheduling for rcu_qs() */
232		resched_cpu(0);
233	}
234	return gp_seq;
235}
236EXPORT_SYMBOL_GPL(start_poll_synchronize_rcu);
237
238/*
239 * Return true if the grace period corresponding to oldstate has completed
240 * and false otherwise.  For more information, see the Tree RCU header
241 * comment.
242 */
243bool poll_state_synchronize_rcu(unsigned long oldstate)
244{
245	return oldstate == RCU_GET_STATE_COMPLETED || READ_ONCE(rcu_ctrlblk.gp_seq) != oldstate;
246}
247EXPORT_SYMBOL_GPL(poll_state_synchronize_rcu);
248
249#ifdef CONFIG_KASAN_GENERIC
250void kvfree_call_rcu(struct rcu_head *head, void *ptr)
251{
252	if (head)
253		kasan_record_aux_stack_noalloc(ptr);
254
255	__kvfree_call_rcu(head, ptr);
256}
257EXPORT_SYMBOL_GPL(kvfree_call_rcu);
258#endif
259
260void __init rcu_init(void)
261{
262	open_softirq(RCU_SOFTIRQ, rcu_process_callbacks);
263	rcu_early_boot_tests();
264	tasks_cblist_init_generic();
265}
266